Early Prediction of Sepsis for ICU Patients using Gradient Boosted Tree

Abstract

Intensive care unit patients, especially those who have undergone surgeries or have severe health issues, tend to have a higher risk of developing sepsis due to a weaker immune system. Due to late detection of sepsis, no preventive actions can be taken to treat sepsis patients. Therefore, this research aims to identify, validate, and test suitable machine learning algorithms for the early prediction of sepsis using pre-processed data produced from the Medical Information Mart for Intensive Care III, MIMIC-III database. This research will be designing prediction models for 15 hours before sepsis onset using pre-processed data obtained from MIMIC-III database using Decision Tree, Random Forest, AdaBoost, Gradient Boosted Tree, and Multilayer Perceptron. A 10 cross-validation is used in validating the models. The performance of prediction models is evaluated mainly using ROC-AUC score. In model comparison, an extra set of prediction models using the same algorithms is developed for 10 hours before sepsis onset to compare its performance with the earlier prediction model developed. The result of model comparison shows that for the prediction model of 15 and 10 hours before sepsis onset, ROC-AUC score for Gradient Boosted Tree is the best with 0.777 for 15 hours and 0.769 respectively from 10 hours prediction model. The results can be optimized further using more data and using derived Boosted Trees algoritms.